So/ar/~nergyvot. 4~, NO. I, pp. I-7, 1990 0038-4192X/00 $3.00+ .

00
Printed in the U.S.A. Copyrisht ~ 1990 Pergamon Press pie

DIFFUSE FRACTION CORRELATIONS

D. T. REINDL, W. A. BECKMAN, and J. A. DUFFLE

Solar Energy Laboratory, University of Wisconsin-Madison, Madison, Wl 53706, U.S.A.

Abstract--The influence of climatic and geometric variables on the hourly diffuse fraction has been studied,
based on a data set with 22,000 hourly measurements from five European and North American locations.
The goal is to determine if other predictor variables, in addition to the clearness index, will significantly
reduce the standard error of Liu- and Jordan-type correlations (ld/l = f(kt)). Stepwise regression is used
to reduce a set of 28 potential predictor variables down to four significant predictors: the clearness index,
solar altitude, ambient temperature, and relative humidity. A piecewise correlation over three ranges of
clearness indices is developed to predict the diffuse fraction as a function of these four variables. A second
piecewise correlation is developed for predicting the diffuse fraction as a function of the clearness index and
solar altitude, for use when temperature and relative humidity are not available. A third piecewisecorrelation
of the Liu- and Jordan-type is developed from the same data set. Comparing this correlation with the new
correlations provides a direct measure of the value of added predictor variables. The full diffuse fraction
correlation reduced the residual sum squares by 14% when compared to the correlation that is a function
of the clearness index only. The correlation including the clearness index and solar altitude diminished the
residual sum squares by 9%. The correlations exhibited some degree of location dependence. This is expected,
as the climates are quite different. The correlations also showed some seasonal dependence; the errors are
higher in the fall and winter than on an annual basis.

1. INTRODUCTION on kt (referred to as Liu- and Jordan-type models) are

convenient because the only required input is hourly
A crucial input required in the transient simulation of global horizontal radiation, I.
solar energy systems is hourly radiation incident on A drawback with using the Liu- and Jordan-type
the collecting surface. Actual measurements of hourly models is the high standard error associated with es-
solar radiation data would be desirable for input but timating the hourly diffuse fraction. The scatter plot
probably are not available for the site and collector of data from Cape Canaveral, FL in Fig. 1 graphically
orientation under consideration. However, hourly illustrates the problem of estimating the hourly diffuse
global radiation on a horizontal surface is one of the fraction as a function of kt only. For example, at kt
most widely available measurements in addition to = 0.5, the measured diffuse fraction ranges from 0.2
other climatic variables (ambient temperature, dew to nearly 1.0. It is clear that the hourly diffuse fraction
point, etc.). Extensive databases exist for a variety of is not a function ofkt only. Models such as Orgill and
locations including SOLMET [ 1], McKay [ 2 ]. Hollands[6] and Erbs[7] provide a single deterministic
If only global horizontal radiation is measured, two value of the hourly diffuse fraction for a given kt. In
problems exist: first, determining the fraction of the an effort to mimic the variation of the diffuse fraction
global which is diffuse (or beam); second, estimating at a particular value of kt, Hollands and Chra [ 11 ] de-
the respective beam, diffuse, and ground reflected veloped a probability density function which allows
components on a tilted surface ofany orientation. The the diffuse fraction to vary about its mean value at a
first of these two problems will be addressed in this given value of kt. Other authors suggest that the vari-
paper. (Both are addressed in [3].) ation of the diffuse fraction for a particular value of kt
Early work by Liu and Jordan[4] showed a rela- is due to other unidentified variables[6,9,10,12,13].
tionship between daily diffuse and daily total radiation Garrison [ 12 ] uses post- 1976 SOLMET data from
on a horizontal surface. Although their original cor- 33 U.S. sites to graphically illustrate the dependence
relation was developed for daily values, it has been ofdiffuse fraction on surface albedo, atmospheric pre-
used for computing the hourly diffuse fraction as a cipitable moisture, atmospheric turbidity, solar ele-
function of the hourly clearness index, kt (ratio of vation, and global horizontal radiation. Without further
hourly global horizontal to hourly extraterrestrial ra- statistical analysis, the relative significance of the vari-
diation) [ 5 ]. Other authors have developed diffuse ables suggested by Garrison remains unknown. Skart-
fraction correlations specifically for hourly intervals. veit and Olseth[13] and Iqbal[9] suggest that the sec-
Orgill and Hollands[6] and Erbs et al.[7] correlate ond most important variable atter kt is the solar alti-
the diffuse fraction with the hourly clearness index. tude.
Stauter and Klein[8] use a clearness index, kc, where This paper focuses on assessing the influence of
a "clear sky" radiation, Ic, replaces extraterrestrial ra- commonly measured climatic variables on the diffuse
diation in the definition of k,. lqbal[9] built on the fraction and correlating the significant variables to re-
work of Bugler[10] to develop a correlation which duce the standard error of Liu- and Jordan-type mod-
predicts hourly diffuse radiation (in the form Ia/lo) as els. First, the data used in this study are introduced.
a function of kt and solar altitude. The models based Second, the influence of commonly measured climatic
2 D.T. REINDL, W . A. BECKMAN, and J. A. DUFFLE

1.0
..-.,o., t . i ~ ,
".: ..
ee "s eoel *' #~% **
0.8
q,-.:..:'
*'%5'~i,.
"~IL
% %e
0.6 e ,e ~
,e i-
L'
*me
~

0.4

0.2

Measured
Cape Data
0.0 ~
0.0 0.2 0.4 0.6 O.O i .0
kt

Fig. I. Measured diffuse fraction vs. clearness index for Cal~ Canaveral, FL.

variables on the diffuse fraction are investigated. Third, pyranometers and pyrheliometers to monitor radiation.
a new hourly diffuse fraction correlation is presented. While the European sites use Kipp solarimeters (shade
Fourth, the relative improvement of the new correla- rings were used to monitor diffuse radiation). An ad-
tion over current Liu- and Jordan-type models is ditional year of data from Oslo, Norway was provided
quantified. by SINTEF and maintained as independent for com-
paring the performance of derived correlations.
The integrity and validity of any empirically derived
2. DATA BASE
correlation relies on the quantity and quality of data
Data from five locations (with a minimum of one used in the model development. It is assumed that a
year of data from each location) is used to investigate sufficient quantity of data exists but the quality of the
the influence of climatic and gcomctric predictors on data need to be examined.
the diffuse fraction. An additional year of data from Three types of datachecks were performed to iden-
an independent location is used to test the derived cor- tify data missing, data which clearly violate physical
relations. Table 1 lists a summary of the sites included limits, and exireme data. When the data were known
in the data base. to be "bad" or "missing," the data fields were filled
The data from Albany, NY were taken under the with a key sequence of numbers (by the reporting lo-
Solar Energy Meteorological Research and Training cation) to clearly indicate the erroneous observation.
Sites (SERMTS) [ 14 ] program at the State University Any hour with data flagged as bad or missing was
of New York, Albany. Data from Cape Canaveral, FL omitted. Second, any hour with an observation that
were provided by the Florida Solar Energy Center violated a physical limit or conservation principle was
(FSEC). (This data set will be referred to as the "Cape'" eliminated from the data set including: reported hours
data.) Three sites with one year of data from each lo- with negative values of radiation, diffuse fraction
cation comprise the European data sets. All sites pro- greater than 1, beam radiation exceeding the extrater-
vide measured values of global horizontal, diffuse (or restrial beam radiation, and dew point temperature
direct normal), ambient temperature, and wet bulb greater than dry bulb temperature. In other circum-
temperature (or dew point). The U.S. sites use Eppley stances, reported data did not exceed physical limits

Table 1. Data base site summary

United States Ettropean
Site Location Albany CapeOmav,~al Co~nha~cn Hamburg Valentia Oslo
Latitude 42.7 north 28.4nor,h 55.7"north 53.$nonh 51.95north 59.56 north
Longitude 73.8 west $0.6wes~ 12.6east I 0.0" cast I0.22 west 10.41" c a ~
StandardMeridian 75,0 west 75.0~west 15.0cast 15.0"cast 0.0 15.0 east
Data Period (From) I/1/79 ]IUSO UU '* I11/" III/* 111/79
Data Pcriod (To) 12/31/82 12/31/80 12/31/* 12/31: 12/31/* 12/31/79

* Year of dataset unknown.

 Diffuse fraction correlations

but were categorized as "extreme" and had to be edited or quantities that can be calculated from commonly
manually. There are various combinations of the dif- observed climatic variables, e.g., ambient temperature,
fuse fraction and clearness index values which would wet bulb temperature, dew point temperature, relative
produce questionable data points. The limits below humidity, etc.
were used to identify such particular cases. (Similar Several steps are necessary to develop an empirical
limits were used by Erbs[7].) Under cloudy overcast model which improves the prediction capabilities of
sky conditions (low values of k,), it is reasonable to the current Liu- and Jordan-type models. The approach
expect that a large portion of the incoming radiation used in this study included the following four basic
will be scattered by the clouds in the atmosphere re- steps: (i) assemble a set of predictor variables; (ii)
sulting in a large diffuse fraction. Case 1 places a limit identify a potential model form; (iii) adopt a predictor
on the diffuse fraction under the cloudy overcast sky selection procedure; and (iv) fit the model.
conditions. If an hour had a measured diffuse fraction Predictor variables are independent variables that
that was less than 0.90 for a clearness index less than may affect the response. It is clear that the response,
0.20, it was eliminated from the data set. Similarly, the diffuse fraction, is influenced by kt. The set of pre-
case 2 places a limit on the diffuse fraction under clear dictors used in this study was limited to kt and other
sky conditions. commonly measured climatic data. Factors such as
atmospheric turbidity and ground reflectance were not
Case 1: I a / I < 0 . 9 0 and kt < 0 . 2 0 included because they are not commonly measured at
radiation monitoring sites. The full set of predictor
Case 2: I d / l > 0.80 and k t > 0.60. variables used in this investigation is given in Table 2.
Monthly average hourly quantities and ratios of the
The quality tests discussed will help eliminate spu- hourly to monthly average hourly values ofthe climatic
rious data and minimize any impact that suspect data predictors and kt were included in an attempt to ac-
would have on a derived correlation. The final data count for possible predictor location dependence. For
set constructed from the measured data that passed all example, the range of ambient temperature over a year
of the quality control checks discussed above produced at Albany will be much larger than the range ofambient
approximately 22,000 hours of data for this study. The temperature for Cape Canaveral, but each location ex-
Osio dataset provided approximately 3000 hours of periences a similar range of measured diffuse fractions
data for independently comparing derived correlations. over the year. However, the ratio of hourly ambient
temperature to monthly average hourly ambient for
both locations may be approximately the same mag-
3. DEVELOPMENT OF NEW HOURLY DIFFUSE nitude. This method of scaling the predictors may im-
FRACTION CORRELATIONS prove the correlations. Exponential terms are included
to represent atmospheric extinction. Geometric terms
The motivation behind investigating hourly diffuse such as solar altitude and optical air mass are included
fraction correlations is to determine if incorporating based on the findings of other authors.
additional predictor variables will significantly reduce With the set of predictors identified, a model form
the standard error of the current Liu- and Jordan-type must be established, e.g., linear, nonlinear, first order,
models. The goal is to find an hourly diffuse fraction second order, etc. The Liu- and Jordan-type models
model which is more accurate than current Liu and are all linear models. Liu and Jordan[5], Orgill and
Jordan type models but remains computationally sim- Hollands [ 6 ], and Erbs [ 7 ] use piecewise fitted models
ple for use in hourly simulation programs such as with varying degree order polynomials. For this study,
TRNSYS [ 5 ]. Also, the inputs to the correlation should a piecewise (in kt) first order linear model of the fol-
be limited to commonly observed climatic variables lowing form will be used to fit the data:

Table 2. Diffuse fraction predictor variables

T~(hr) T.~(hr)
~(hr) ~(hr)
T.(hr)l~.(hr) T,,b ( hr )ff'~b ( hr )
T~,(hr) k,(hr)
"I~(hr) ~(hr)
T#(hr)lT~(hr) k,(hr)l~(hr)
t.o(hr ) t~(hr)
~o(hr ) "~(hr)
o~(hr )f'o~(hr ) t~(hr )~( hr )
sin(a) T,(hr). t~(hr)
m = 1 /cos(0,) t~(hr) / [7",+ 273]
k,(hr) . m tp(hr)~(hr). T,,(hr)
exp [sin(a)] exp [t~(hr)l(To(hr))]
exp [k, .m] exp [t)(hr)/(~(hr). T(hr))]
 D. T. REINDL,W. A. BECKMAN,and J. A. DUFFLE

Y = ~o + 8~x~ + ~:x2 + " " + ~jxj + ~ (1) imize the standard error of the final correlation. A
manual search technique was employed to find the
where y is the true response, Bj is thej th model param- clearness index interval which minimized the standard
eter, xj is t h e j th predictor, and ~ represents the model error of the correlation. The center interval which
error. minimized the standard error is 0.3 < kt < 0.78.
In an effort to gain understanding of the diffuse The standard error for the correlation in the center
fraction's association with each predictor variable, interval is 0.129. The standard error for the correlation
analysis was performed on the center interval of kt which included the monthly average predictors is also
(initially assuming the same interval as Orgill and 0.129 (based on the interval 0.35 < kt < 0.75). Thus,
Hollands, 0.35 < kt < 0.75). Stepwise regressions were it appears that there is not a significant loss (from an
performed on a monthly basis for each location to de- error standpoint) by not including the monthly average
termine if location or seasonal bias existed in the se- hourly predictor variables in the correlation.
lection of the best set of predictors. (For information The clearness index is the most important variable
regarding stepwise regression, see Draper and in the low and middle intervals but at the high interval,
Smith[15 ].) The complete set of predictor variables the significance of the clearness index decreases dra-
listed in Table 2 was used in the stepwise procedure. matically. The solar altitude effects are not as strong
On a monthly basis, there was not a great deal ofcon- under cloudy skies (low values of k,) but under clear
sistency in the variable selection due to the relatively skies (high values of k,), the solar altitude becomes
short time interval; therefore, stepwise regression was the dominant predictor variable. For clear sky condi-
performed on a yearly basis for each location. The four tions, the diffuse fraction increases for decreasing solar
best predictor variables to explain the deviations in the altitude angles due to the longer path length required
diffuse fraction are kt, sin(a), 7~a, /~. for radiation to travel. These results are consistent with
Unfortunately, monthly average hourly data are not those found by Skartveit and Olseth [13].
included in commonly used data sets[l,2]. A corre- The final version of the full correlation is given
lation with monthly average hourly variables would below.
force users to derive the necessary quantities from ex-
isting data sets. At this point, a decision was made to Interval: 0 ~ kt g 0.3; Constraint: I d / l < 1.0.
eliminate the use of monthly average hourly predictors
I d / l = 1.000 -- 0.232kt + 0.0239 sin(a)
and pursue a correlation based only on hourly values
of the predictors. The impact of this decision will be - 0.000682Ta + 0.01950 (2a)
investigated when the new set of predictor variables
are selected. Interval: 0.3 < k, < 0.78;
The stepwise selection procedure was applied to the Constraint: Id/I<0.97 and I d / l > O.l.
center interval of kt using the set of predictors with the
monthly average hourly predictors removed. On an Id/I -- 1.329 - 1.716k, + 0.267 sin(a)
hourly basis, the top four predictors were selected; the - 0.00357To + 0.106~ (2b)
best predictors to explain the deviations in the diffuse
fraction are kt, sin(a), Ta, ~. Interval: 0.78 < kt; Constraint: I~/I > 0.1.
This set of predictors provides the foundation for Id/I = 0.426kt - 0.256 s i n ( a )
the remaining development and analysis of a new dif-
fuse fraction correlation.* As will be shown, substi- + 0.00349Ta + 0.0734~. (2c)
tuting Ta for 7"o and 0 for ~ / ~ does not significantly
effect the results. Because the above piecewise correlation includes
The next step in the diffuse fraction correlation de- multiple predictor variables, it is possible that some
velopment is to determine the best kt intervals for combinations of predictors may produce unreasonable
piecewise fitting. The appropriate kt interval will rain- values of the diffuse fraction, e.g., greater than 1; there-
fore, subsequent constraints are placed on the corre-
lation in each interval to assure reasonable predicted
* The reviewers suggestedthat precipitable water may be values. The constraint values on the middle and high
an important predictorvariableand should have been included
in Table 2. Precipitable water is not a commonly measured interval of k, are based on the observations from the
variable but several correlations of precipitable water vapor data sets.
in terms of dew point temperature[16,17] are available. The At this point, a simple piecewise correlation exists
Bolsenga[ 16] and Smith[ 17] correlationsfor precipitablewater which gives the hourly diffuse fraction as a function
vapor were included in a subset of predictor variables (the of hourly clearness index, solar altitude, ambient tem-
subset includedkt, sin(a), :ira,O, and precipitablewater vapor)
and the stepwiseselectionprocedure was rela~ated.Precipitable perature, and relative humidity. It would be desirable
water vapor was not selected as one ofthe top four predictors to provide a reduced form of the current correlation
in the interval 0 < kt < 0.78. In the interval k~> 0.78, where for use when hourly ambient temperature a n d / o r rel-
only 3% of the data lie,precipitablewater vapor was the second ative humidity data are not available. The result is a
most important predictorvariable.A correlationthat explicitly
includesk. precipitablewater, and ambient temperature was pieccwise model which provides estimates oftbe hourly
derived but the standard error was slightly greater than the diffuse fraction as a function ofthe clearness index and
correlation derived based on the above four variables. solar altitude angle.
 Diffuse fraction correlations

Interval: 0 < kt < 0.3; Constraint: I a / I < 1.0. tion correlation (eqns 2) and the current Liu- and Jor-
dan-type models, the correlation based on k, (eqns 4)
I a / l = 1.020 - 0.254k, + 0.0123 s i n ( a ) (3a)
derived from the existing data set will be included in
Interval: 0.3 < kt < 0.78; all model comparisons. By comparing the ktcorr model
with the new diffuse fraction correlation, the relative
Constraint: I e / I < 0.97 and l a / l > O.l.
merit of added climatic and geometric terms in the
l, d l = 1.400 - 1.749kt + 0.177 s i n ( a ) (3b) new model will be directly assessed. The reduced cor-
relation based on the clearness index and solar altitude
Interval: 0.78 < kt; Constraint: l a / l > O. 1. (eqns 3) will also be included in the correlation com-
I d / I = 0.486kt - 0.182 s i n ( a ) . (3c) parison. The only other existing model that will be
included in the model comparisons will be the Erbs [ 7 ]
A final correlation which is a function of kt only model. The Orgill and Hoilands model is similar to
was also developed. This correlation (identified as the Erbs models so that conclusions about the Erbs
"ktcorr") will allow direct comparison of the new cor- model also apply to the Orgill and HoUands model.
relations to the Liu- and Jordan-type correlations. The Location and seasonal effects are noted. The CRSS is
correlation, ktcorr, is given below. calculated by the following relationship:

Interval: 0 < kt < 0.3; Constraint: I ~ / I < 1.0. CRSS = z[(I~/t)p~ - (ls//)m~]:. (5)
I,t/I = 1.020 - 0.248kt (4a) Applying eqn (5) to the complete data set yielded
the results listed in Table 3. On an overall basis, the
Interval: 0.3 < k, < 0.78
new hourly diffuse fraction correlation reduces the re-
I a / I = 1 . 4 5 - !.67k, (4b) sidual sum squares of the correlation based on kt only
(derived from the same data set) by 14.4%. The reduced
Interval: 0.78 < k,
hourly diffuse fraction correlation shows a 9% im-
I d / l = .147. (4C) provement over ktcorr. Included in Table 3 are results
for one year of independent data from Oslo, Norway.
The derived correlation based on k, is similar to Orgill The correlations given by eqns (2) and (3) showed
and Hollands[6] and Erbs[7] as shown in Fig. 2. (Each 17% and 13.5% reduction in residual sum squares when
of these three correlations was based on an entirely compared with ktcorr. An improvement of 26% and
independent data set.) 23% is gained when compared with the Erbs' corre-
lation at the same location.
The derived correlations exhibit slight location and
4. MODEL PERFORMANCE
seasonal dependencies. Locatiou differences are noted
A simple composite residual sum squares (CRSS) in Fig. 3 by plotting the residual mean squares ( R S S /
comparison is used to quantify the improvement of (n - p)) for each location. The variation in the residual
the new hourly diffuse fraction correlation over current mean squares for each location suggests that the cor-
Liu- and Jordan-type models. In an effort to provide relations are not entirely location independent. Co-
a fair comparison between the new hourly diffuse frac- penhagen exhibits the largest residual mean square.

1.0

0.8

0.6

0.4.

0.2
: : : : : Orglll and Hollands
ErBs
c = = = : k t ~ o r r

0.0 I I I I I I I I I I

0.0 0.2 0.4 0.6 0.8 1.0

kt
Fig. 2. Diffuse fraction cor~ladons including lacom
 D. T. REINDL, W. A. BECKMAN,and J. A. DUFFLE

Table 3. Composite residual sum squares results

Location n Eqn. 2 Eqn. 3 Eqn. 4 Erbs

Albany, '79 2178 33.04 35.67 38.16 37.01
Albany, '80 2214 35.42 36.05 39.64 37.73
Albany, '81 2126 27.37 29.72 32.96 30.99
Albany, '82 1994 23.38 25.07 28.80 29.54
Cape 3596 45.85 50.63 55.29 57.97
Valcntia 3386 37.95 41.04 46,37 47.24
Hamburg 3279 34.75 34.51 35.59 37.30
Copenhagen 3150 55.49 58.52 65.78 77.64
Total 21923 293.25 311.21 342.59 355.42
Oslo = 2927 35.27 36.76 42.52 47.74

Oslo dataset was not used in deriving the correlations given by eqns (2)-(4).

However, the overall variation in residual mean square the fall and winter m o n t h s than on an annual basis.
from location to location is o f the same order as the No a t t e m p t is m a d e to account for location or seasonal
yearly variation in residual mean square for Albany, effects. The authors feel that the current correlation is
1979-1982. Also, residual mean squares are higher in acceptable.

2.5

2 -

o 1.5
O

r-
v

0~
n"
1

0.5

Ab7g AIbS0 Ab81 AIb82 Cape Vabn Hamb Copen

Location

Full ~ Reduced ~-~ ktcorr

Fig. 3. Location residual mean squares.
 Diffuse fraction correlations 7

5. CONCLUSIONS relative humidity (fraction)

humidity ratio
The goal of this research was to reduce the standard
Subscripts
error of the current Liu- and Jordan-type correlations
by including additional predictor variables in the i datum
model. Stepwise regression is used to reduce a large set meas measured data
pred model predicted
of potential predictor variables down to four significant
predictors. The significant predictors include hourly
values of clearness index, solar altitude, ambient tem-
REFERENCES
perature, and relative humidity. The final version of
the piecewise correlation is given by eqns (2). In the I. SOLMET, Volume l--User's Manual, Hourly solar ra-
event that hourly ambient temperature a n d / o r hourly diation surface meteorological observations. #TD-9724
relative humidity data are not available, a reduced form (1978).
of the correlation was derived to predict the diffuse 2. D. C. McKay et aL, As noted in Final Report lEA Task
IX "Validations of models for estimating solar radiation
fraction as a function of clearness index and solar al-
on horizontal surface.'" Data requests referred to Dr.
titude (eqns (3)). D. C. McKay, Canada Climate Centre, The Atmospheric
The new correlation reduced the composite residual Environment Service, 4905 Dufferin Street, Downsview,
sum of squares by 14.4% when compared to a k, cor- Ontario, M3H5T4, Canada.
relation derived from the same data set. The reduced 3. D.T. Reindl, Estimating diffuse radiation on horizontal
surfaces and total radiation on tilted surface. MS Thesis,
form of the correlation reduces the composite residual University of Wisconsin-Madison, Madison, W1 (1988).
sum squares by 9.2%. When an independent data set 4. B.Y.H. Liu and R. C. Jordan, The interrelationship and
is used, the new correlation reduced the residual sum characteristic distribution of direct, diffuse, and total solar
of squares by 26% compared to the Erbs' correlation. radiation, Solar Energy 4, 1-19 (1960).
5. S. A. Klein et al.. "TRNSYS--A transient simulation
Some location and seasonal dependencies were sug-
program," University of Wisconsin-Madison, Engineer-
gested but their effects are considered to be negligible ing Experiment Station Report 38-12, Version 12.2
to the correlation's overall performance. ( 1988 ).
6. J. F. Orgill and K. G. T. Hollands, Correlation equation
for hourly diffuse radiation on a horizontal surface, Solar
Acknowledgments--This research was funded by the United Energy 19, 357 (1977).
States Department of Energy. A special thanks to Ron Stewart 7. D.G. Erbs, Methods for estimating the diffuse fraction of
at State University of New York, Albany; Safvat Kalaghchy hourly, daily and monthly-averageglobal solar radiation.
at the Florida Solar Energy Center; and Atle Nordgaard from MA Thesis, Mechanical Engineering University of Wis-
SINTEF Trondheim, Norway for providing data. consin-Madison, Madison, Wl (1980).
8. J. A. Duffle and W. A. Beckman, Solar engineering of
thermal processes, Wiley-lnterseience Publication, New
NOMENCLATURE York, pp. 1, 72 (1980).
9. M. lqbal, Prediction of hourly diffuse solar radiation from
measured hourly global radiation on a horizontal surface,
CRSS Composite residual sum squares for all locations Solar Energy 24, 491-503 (1980).
RSS Composite residual sum squares for one location 10. J. M. Bugler, The determination of houdy insolation on
sc solar constant (a value of 1353 W / m z was used) an inclined plane using a diffuse irradiance model based
hr hourly on hourly measured global horizontal insolation, Solar
1 hourly total radiation on a horizontal surface, kJ/m z Energy 19, 477--491 (1977).
hourly beam radiation on a horizontal surface, kJ/ 1!. K. G. T. Hollands and S. J. Chra, A probability density
m 2 function for the diffuse fraction, with applications, Solar
hourly diffuse radiation on a horizontal surface, kJ/ Energy 38, 237-245 (1987).
m 2 12. J. D. Garrison, A study of the division ofgiobal irradiance
1o hourly extraterrestrial radiation on a horizontal sur- into direct and diffuse irradiance at thirty three U.S. sites,
face, kJ/m 2 Solar Energy 35, 341-351 ( 1985 ).
k, hourly clearness index, I/1o 13. A. Skartveit and J. A. Olseth, A model for the diffuse
m optical air mass fraction of hourly global radiation, Solar Energy.38, 271-
n number of observations 274 ( 1987 ).
p number of parameters in the model 14. Solar Energy Meteorological Research and Training Site
RSS Composite residual sum squares for one location Program (SEMRTS), operated for U.S. Department of
SS Sum of squares Energy by Meteorological Research Institute under con-
To ambient temperature, C tract No. EG-77-C-01-4042.
T~, dew point temperature, *C 15. N. R. Draper and H. Smith, Applied regression analysis,
T~ wet bulb temperature, *C Wiley, New York ( 1981 ).
g monthly average hourly value o f x 16. S.J. Bolsenga, The relationship between total atmospheric
water vapor and surface dew point on a mean daily and
Greek hourly basis, Journal of Applied Meteorology 4, 430-432
(1965).
Ol solar altitude angle 17. W. L. Smith, Note on the relationship between total pre-
Be true model parameter cipitable water and surface dew point, Journal of Applied
E model error, residual Meteorology 5, 726-727 (1966).